Prediction accuracy and stability of regression with optimal
scaling transformations
Kooij, A.J. van der
Citation
Kooij, A. J. van der. (2007, June 27). Prediction accuracy and stability of regression with optimal scaling transformations. Leiden. Retrieved from https://hdl.handle.net/1887/12096
Version: Corrected Publisher’s Version
License: Licence agreement concerning inclusion of doctoral thesis in the Institutional Repository of the University of Leiden
Downloaded from: https://hdl.handle.net/1887/12096
Note: To cite this publication please use the final published version (if applicable).
Prediction Accuracy and Stability
of Regression with
Optimal Scaling Transformations
Van der Kooij, Anita Jolande,
Prediction Accuracy and Stability of Regression with Optimal Scaling Transformations.
Dissertation Leiden University — With ref. — With Summary in Dutch.
Subject headings: nonlinear regression; CATREG; optimal scaling;
transformations; local minima; prediction accuracy; regularization;
.632 Bootstrap; Ridge regression; Lasso; Elastic Net ISBN 978-90-9021936-3
2007 Anita J. van der Kooijc
Printed by Mostert en van Onderen, Leiden
Prediction Accuracy and Stability of
Regression with Optimal Scaling
Transformations
Proefschrift ter verkrijging van
de graad van Doctor aan de Universiteit Leiden,
op gezag van de Rector Magnificus prof.mr. P.F. van der Heijden, volgens besluit van het College voor Promoties
te verdedigen op woensdag 27 juni 2007 klokke 16.15 uur
door
Anita J. van der Kooij geboren te Boskoop
in 1961
PROMOTIECOMMISSIE
Promotor Prof. dr. J.J. Meulman
Referent Prof. J.H. Friedman, Ph.D., Stanford University, USA Overige Leden Prof. dr. W.J. Heiser
Prof. dr. M.H. van IJzendoorn Dr. ing. P.H.C. Eilers
Prof. dr. R.D. Gill
Contents
1 Introduction 1
1.1 Regression methods for nonlinearly related data . . . 2
1.2 The CATREG method . . . 3
1.2.1 Optimal scaling levels . . . 5
1.2.2 Estimation of transformations . . . 8
1.2.3 Local minima . . . 10
1.2.4 Prediction accuracy, stability, and regularization . . . . 11
1.3 Outline . . . 12
2 Local Minima in Regression with Optimal Scaling Transfor- mations 15 2.1 Introduction . . . 16
2.2 Model . . . 17
2.3 Algorithm . . . 18
2.4 Local minima . . . 21
2.4.1 A probable cause of local minima . . . 21
2.4.2 Strategies to obtain the global minimum . . . 23
2.5 Simulation study . . . 26
2.5.1 Design . . . 26
2.5.2 Results . . . 28
2.6 Conclusion . . . 35
3 Prediction Accuracy of Regression with Optimal Scaling trans- formations: The .632 Bootstrap with CATREG 37 3.1 Introduction . . . 37
3.2 CATREG: Regression with Optimal Scaling Transformations . 40 3.2.1 CATREG model . . . 40
3.2.2 CATREG algorithm . . . 41
3.3 Estimation of expected prediction error . . . 44 v
vi CONTENTS
3.3.1 The .632 bootstrap with linear regression . . . 45
3.3.2 The .632 bootstrap with CATREG . . . 46
3.4 Performance of CATREG and six other regression with trans- formation methods: Prediction accuracy in the analysis of the Ozone data . . . 47
3.4.1 Comparison between CATREG and Box-Tidwell Method, Full Additive Model, Stepwise Full Additive Model, TURBO, and BRUTO . . . 48
3.4.2 Comparison between CATREG and ACE . . . 50
3.5 Prediction accuracy for different scaling levels in CATREG . . 53
3.6 The effect of the number of observations on the expected pre- diction error . . . 58
3.7 Conclusions . . . 61
3.8 Computational note . . . 63
4 Regularization with Ridge penalties, the Lasso, and the Elas- tic Net for Regression with Optimal Scaling Transformations 65 4.1 Introduction . . . 66
4.2 Ridge penalties, the Lasso, and the Elastic Net for linear re- gression . . . 67
4.3 Ridge penalties, the Lasso, and the Elastic Net with CATREG 70 4.3.1 Updating the regularization regression coefficients in CATREG . . . 71
4.3.2 Paths for the coefficients . . . 72
4.3.3 Finding the transition points in CATREG . . . 74
4.3.4 Including nonlinear transformations in the CATREG al- gorithm . . . 75
4.4 Selection of the optimal penalty parameter . . . 75
4.4.1 Illustration . . . 76
4.5 Shrinking a nominal variable versus shrinking it’s categories . . 80
4.6 Discussion . . . 86
5 Application of the CATREG-Lasso: Severity of Bulimia Ner- vosa - Components of the Syndrome and Definition of Ther- apy Outcome 91 5.1 Introduction . . . 92
5.2 Method . . . 94
5.2.1 Sample and instruments . . . 94
5.2.2 Self and Expert Ratings . . . 97
5.2.3 Statistical Analysis . . . 97
CONTENTS vii
5.3 Results . . . 101
5.3.1 Model selection . . . 101
5.3.2 Discrimination between health and pathology . . . 104
5.4 Discussion . . . 105
6 General Discussion 109 6.1 A short retrospect . . . 109
6.2 Topics for further research . . . 114
6.2.1 Transformations towards independence and regularization114 6.2.2 The efffect of shrinking on low-frequency categories . . . 114
6.2.3 Regularization in Discriminant Analysis . . . 115
6.2.4 CATREG with cluster restrictions . . . 115
Appendix A CATREG Algorithm 117 Appendix B CATREG sections from SPSS Categories R 11.0 131 Appendix C Notation 179 References 181 Summary in Dutch (Samenvatting) 189 Curriculum vitae 197 Overview of Applications CATREG for Diamonds data . . . 6
Prediction accuracy for Ozone data . . . 47
Effect of number of observations on prediction accuracy for Demographic data . . . 58
Linear Lasso for Diabetes data . . . 73
Linear and nonlinear Ridge, Lasso, and Elastic Net for Prostate cancer data . . . 76
Nonlinear Lasso and dummies-Lasso for Breast cancer data . . . 82
Nonlinear Lasso and .632 bootstrap for Bulimia Nervosa data . . . . 91